IVIM in the Body: A General Overview

Understanding diffusion‐weighted MRI analysis: Repeatability and performance of diffusion models in a benign breast lesion cohort

“…It has been demonstrated that repeatability of the ADC in extracranial oncological applications can be generalised, 16 but similar work for non-Gaussian diffusion analyses remains to be fully explored. 20 The main limitation of this study is that it reflects only a small cohort of benign breast lesions, some of which were in themselves small, where nonmonoexponential signal behaviour has been shown to be less prominent than for malignant lesions that are of greater clinical interest. 44,45 Nevertheless, the signal curves are clearly shown to be nonmonoexponential, and to dismiss the utility of a benign lesion cohort on the basis that the pseudodiffusion component is smaller illustrates precisely the problem with beginning from an unstated assumption of the correctness of one particular model (in this case IVIM).…”

“…The repeatability thus explicitly represents elements from underlying physiology and from patient set-up and scanner performance, to most accurately reflect clinical use. DWI acquisition, acquired prior to any contrast administration, used a spin-echo echo planar imaging (SE-EPI) sequence, in unilateral sagittal orientation, with parameters including: bipolar diffusion encoding, TR/TE: 11,600/86 ms, FOV: 180 x 180 mm, matrix: 90 x 90, slice thickness: 2.5 mm, slices: 60, iPAT: GRAPPA 2, spectral attenuated inversion recovery [SPAIR]), and 13 b-values: 0, 10,20,30,40,50,70,90,120,150,200, 400 and 700 s.mm −2 , thereby placing an emphasis on fine sampling of the lowest b-values and retaining higher signal at the maximum diffusion-weighting. Diffusion times were Δ = 40.8 ms and δ = 19.5 ms. An additional phase-reversed, geometry-matched scan without diffusionweighting (b = 0 s.mm −2 ) was acquired to allow for correction of geometric distortion arising from susceptibility boundaries.…”

“…Repeat-measure CoV and 95% limits of agreement (LoA) were calculated for all DWI parameters for the values at every percentile across the VOI histogram for voxel-wise results, and for the mean and median VOI values. 17,20 Median diffusion parameters were compared using a student's t-test, with a significance threshold of 0.05. All statistical analyses were performed in MATLAB using standard functions.…”

Accuracy of breast cancer lesion classification using intravoxel incoherent motion diffusion‐weighted imaging is improved by the inclusion of global or local prior knowledge with bayesian methods

“…The IVIM model is thus formulated as the sum of two exponential components, of empirical coefficients D* and D , and the volume fraction f quantifying their relative contributions to the total signal. While IVIM has potential for noninvasive perfusion measurement, the method has not been widely clinically applied due to high noise sensitivity and lack of standardization, with studies demonstrating significantly different IVIM parameter values for the same tissue type and in similar patient groups; additionally, pseudodiffusion parameters f and D* are generally less repeatable than D . Alongside the data acquisition strategy, image quality, and physiological noise, IVIM parameters may depend to some degree on the signal fitting methods/algorithms used …”

“…The IVIM model is thus formulated as the sum of two exponential components, of empirical coefficients D* and D, and the volume fraction f quantifying their relative contributions to the total signal. While IVIM has potential for noninvasive perfusion measurement, the method has not been widely clinically applied due to high noise sensitivity and lack of standardization, with studies demonstrating significantly different IVIM parameter values for the same tissue type 2 and in similar patient groups 3 ; additionally, pseudodiffusion parameters f and D* are generally less repeatable than D. 4,5 Alongside the data acquisition strategy, image quality, and physiological noise, IVIM parameters may depend to some degree on the signal fitting methods/algorithms used. 2,3,[6][7][8][9][10][11][12] The most common method is to use some form of nonlinear least squares (LSQ) fitting, which may be constrained or unconstrained to specified limits for each parameter; however, the noise sensitivity of this approach has led to exploration of alternatives that attempt to fit individual components separately in a segmented approach, 3,[13][14][15][16] or to include some degree of prior knowledge, such as expectations of either local or global homogeneity, in more complex Bayesian-type approaches that are able to generate much cleaner parameter maps compared with those from nonlinear least squares.…”